Entropy-Guided Robust Feature Domain Adaptation for Eeg-Based Cross-Dataset Drowsiness Recognition

Driver drowsiness is one of the main factors that cause road accidents. Development of drowsiness monitoring system is of top priority for road safety and accident prevention. By monitoring voltage fluctuations on the scalp that reflect mental activities taking place in the brain, electroencephalography (EEG) has been regarded as one of the best methods to detect drowsiness. However, the main problem of EEG is its low signal-to-noise rate and vulnerability to various kinds of noise that cause both individual-level variations and group-level drifts due to different devices and environments. In order to solve the problem, we propose an entropy-guided robust feature (EGRF) adaptation framework, which uses the state-of-the-art model named Interpretable Convolutional Neural Network (ICNN) as backbone to extract shared features across different subjects and a novel unsupervised domain adaptation (UDA) technique to minimize the group-level drifts of EEG data. We use two public driving datasets SEED-VIG and SADT to test the method on the cross-dataset setting. Results show that the model has achieved a mean accuracy of 75.1% for two-class drowsiness recognition on 11 subjects when SADT is used as the source dataset and Seed-Vig is used as target, which is higher than the baseline methods ranging from 60.1% to 68.6%. On the reverse setting, the model achieves a mean accuracy of 80% on 12 subjects, which is higher than baseline from ranging from 60.4% to 73.1%. Our work illustrates a promising direction of using EEG for calibration-free driver drowsiness recognition system.